YOLO Vending Machine Inventory, Vandalism & Loitering Detection on NPU Edge Hardware
A vending machine operator needed a fully offline system to detect empty slots, identify door tampering, and flag loitering in real time — without a cloud dependency. We trained YOLOv11 on a custom vending interior and exterior dataset and deployed it to NPU-equipped edge hardware using INT8 quantisation. The unit processes its own camera feed at 60fps on-device, buffers alerts locally, and syncs clips and logs when connectivity resumes.
Discuss a Similar ProjectWhat We Built
Custom YOLOv11 on Vending Interior Dataset
Trained from scratch on annotated images of vending product slots across 40 machine SKU types — occupied, empty, partially filled, and obstructed states — with per-slot grid mapping for restocking dispatch accuracy.
Slot Occupancy Detector with Restocking Alerts
Per-column, per-row slot occupancy tracked in real time. When any product column drops below the configured threshold, a restocking alert is queued with machine ID, slot coordinates, and item SKU associated to that position.
Vandalism & Tamper Classifier
Separate YOLO head monitors door status, bezel integrity, and impact patterns. Distinguishes legitimate door access (staff keycard events) from forced-entry attempts — triggering an immediate priority alert with timestamped clip.
Loitering Tracker with Configurable Dwell Time
Person-tracking algorithm assigns persistent IDs to individuals in the camera zone. Dwell time threshold is configurable per-machine (default 90 seconds). Loitering events recorded with first-seen, last-seen timestamps and thumbnail grid.
NPU-Optimised INT8 Inference
Model converted to ONNX and optimised for NPU execution via Rockchip, Hailo, or Axera SDK depending on hardware variant. INT8 quantisation reduces model memory footprint to under 14MB while preserving detection accuracy above 94% mAP.
Edge-Cloud Sync on Connectivity Resume
All events buffered locally over MQTT with persistent SQLite queue. When the device reconnects (Wi-Fi, 4G, or LAN), the sync daemon uploads alert clips, slot snapshots, and structured log batches to the central management dashboard via FastAPI.
Technologies Used
Key Outcomes
Real-time inference on NPU without fan, without cloud, at room temperature
Precision on loitering detection and tamper classification in production
Average restocking dispatch compared to scheduled manual checks
Need Something Similar?
We work with teams deploying computer vision to physical hardware in spaces with unreliable or zero connectivity. Tell us your use case and target device.